mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-05-13 03:52:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 8

fix(gateway): scrub memory-context leaks from vision auto-analysis output

Browse source 
fixes #5719

The auxiliary vision LLM called by gateway._enrich_message_with_vision
can echo its injected Honcho system prompt back into the image
description.  That description gets embedded verbatim into the enriched
user message, so recalled memory (personal facts, dialectic output)
surfaces into a user-visible bubble.

Strips both forms of leak before embedding:
  - <memory-context>...</memory-context> fenced blocks (sanitize_context)
  - trailing '## Honcho Context' sections (header + everything after)

Plus regression tests:
  - tests/agent/test_streaming_context_scrubber.py — 13 tests on the
    stateful scrubber (whole block, split tags, false-positive partial
    tags, unterminated span, reset, case-insensitivity)
  - tests/run_agent/test_run_agent_codex_responses.py — 2 new tests on
    _fire_stream_delta covering the realistic 7-chunk leak scenario and
    the cross-turn scrubber reset
  - tests/gateway/test_vision_memory_leak.py — 4 tests covering the
    vision auto-analysis boundary (clean pass-through, '## Honcho Context'
    header, fenced block, both patterns together)
This commit is contained in:
Erosika 2026-04-24 18:33:19 -04:00
parent 6fe522cabd
commit edc23e888d
4 changed files with 309 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

9
gateway/run.py
View file

@ -8253,6 +8253,7 @@ class GatewayRunner:
The enriched message string with vision descriptions prepended.
"""
from tools.vision_tools import vision_analyze_tool
from agent.memory_manager import sanitize_context
analysis_prompt = (
"Describe everything visible in this image in thorough detail. "
@ -8271,6 +8272,14 @@ class GatewayRunner:
result = json.loads(result_json)
if result.get("success"):
description = result.get("analysis", "")
# The auxiliary vision LLM can echo injected system-prompt
# memory context back into its output (#5719). Scrub any
# <memory-context> fences and the "## Honcho Context"
# section before the description lands in a user-visible
# message.
description = sanitize_context(description)
if "## Honcho Context" in description:
description = description.split("## Honcho Context", 1)[0].rstrip()
enriched_parts.append(
f"[The user sent an image~ Here's what I can see:\n{description}]\n"
f"[If you need a closer look, use vision_analyze with "